Download PPTX

MapReduce
Simplified Data Processing on Large Clusters
(Without the Agonizing Pain)
Presented by Aaron Nathan
The Problem
• Massive amounts of data
– >100TB (the internet)
– Needs simple processing
• Computers aren’t perfect
– Slow
– Unreliable
– Misconfigured
• Requires complex (i.e. bug prone) code
MapReduce to the Rescue!
• Common Functional Programming Model
– Map Step
map (in_key, in_value) -> list(out_key, intermediate_value)
• Split a problem into a lot of smaller subproblems
– Reduce Step
reduce (out_key, list(intermediate_value)) -> list(out_value)
• Combine the outputs of the subproblems to give the
original problem’s answer
• Each “function ” is independent
• Highly Parallelizable
Algorithm Picture
DATA
MAP
MAP
K1:v K1:v K2:v
K2:v
MAP
K1:v
MAP
MAP
K2:v K3:v
K3:v
aggregator
K1:v,v,v
K2:v,v,v
K3:v,v,v
REDUCE
REDUCE
REDUCE
Answer!
Answer!
Answer!
Some Example Code
map(String input_key, String input_value):
// input_key: document name
// input_value: document contents
for each word w in input_value:
EmitIntermediate(w, "1");
reduce(String output_key, Iterator intermediate_values):
// output_key: a word
// output_values: a list of counts int result = 0;
for each v in intermediate_values:
result += ParseInt(v);
Emit(AsString(result));
Some Example Applications
•
•
•
•
•
•
Distributed Grep
URL Access Frequency Counter
Reverse Web Link Graph
Term-Vector per Host
Distributed Sort
Inverted Index
The Implementation
• Google Clusters
– 100s-1000s Dual Core x86 Commodity Machines
– Commodity Networking (100mbps/1Gbps)
– GFS
• Google Job Scheduler
• Library linked in c++
Execution
The Master
• Maintains the state and identify of all workers
• Manages intermediate values
• Receives signals from Map workers upon
completion
• Broadcasts signals to Reduce workers as they
work
• Can retask completed Map workers to Reduce
workers.
In Case of Failure
• Periodic Pings from Master->Workers
– On failure resets state of assigned task of dead worker
• Simple system proves resilient
– Works in case of a 80 simultaneous machine failures!
• Master failure is unhandled.
• Worker Failure doesn’t effect output
(output identical whether failure occurs or not)
– Each map writes to local disk only
– If a mapper is lost, the data is just reprocessed
– Non-deterministic map functions aren’t guaranteed
Preserving Bandwidth
• Machines are in racks with small interconnects
– Use location information from GFS
– Attempts to put tasks for workers and input slices
on the same rack
– Usually results in LOCAL reads!
Backup Execution Tasks
• What if one machine is slow?
• Can delay the completion of the entire MR
Operation!
• Answer: Backup (Redundant) Executions
– Whoever finishes first completes the task!
– Enabled towards the end of processing
Partitioning
• M = number of Map Tasks
(the number of input splits)
• R = number of Reduce Tasks
(the number of intermediate key splits)
• W = number of worker computers
• In General:
– M = sizeof(Input)/64 MB
– R = W*n (where n is a small number)
• Typical Scenario:
InputSize = 12 TB,
M = 200,000,
R = 5000
W = 2000
Custom Partitioning
• Default Partitioned on intermediate key
– Hash(intermediate_key) mod R
• What if user has apriori knowledge about the
key?
– Allow for user-defined hashing function
– Ex. Hash(Hostname(url_key))
The Combiner
• If reducer is associative and communivitive
– (2 + 5) + 4 = 11 or 2 + (5 + 4) = 11
– (15+x)+2=2+(15+x)
• Repeated intermediate keys can be merged
– Saves network bandwidth
– Essentially like a local reduce task
I/O Abstractions
• How to get initial key value pairs to map?
– Define an input “format”
– Make sure splits occur in reasonable places
– Ex: Text
• Each line is a key/pair
• Can come from GFS, bigTable, or anywhere really!
– Output works analogously
Skipping Bad Records
• What if a user makes a mistake in map/reduce
• And only apparent on few jobs..
• Worker sends message to Master
• Skip record on >1 worker failure and tell
others to ignore this record
Removing Unnecessary Development Pain
• Local MapReduce Implementation that runs
on development machine
• Master has HTTP page with status of entire
operation
– Shows “bad records”
• Provide a Counter Facility
– Master aggregates “counts” and displayed on
Master HTTP page
A look at the UI (in 1994)
http://labs.google.com/papers/mapreduce-osdi04-slides/index-auto-0013.html
Performance Benchmarks
Sorting
AND
Searching
Search (Grep)
• Scan through 1010
100 byte records
(1TB)
• M = 15000, R = 1
• Startup time
– GFS Localization
– Program Propagation
• Peak -> 30 GB/sec!
Sort
•
•
•
•
50 lines of code
Map->key + textline
Reduce-> Identity
M = 15000, R = 4000
– Partition on init bytes of
intermediate key
• Sorts in 891 sec!
What about Backup Tasks?
And wait…it’s useful!
NB: August 2004
Open Source Implementation
• Hadoop
• http://hadoop.apache.org/core/
– Relies on HDFS
– All interfaces look almost exactly like MapReduce
paper
• There is even a talk about it today!
– 4:15 B17 CS Colloquium:
Mike Cafarella (Uwash)
Active Disks for Large-Scale
Data Processing
The Concept
• Use aggregate processing power
– Networked disks allow for higher throughput
• Why not move part of the application onto
the disk device?
– Reduce data traffic
– Increase parallelism further
Shrinking Support Hardware…
Example Applications
• Media Database
– Find similar media data by “fingerprint”
• Real Time Applications
– Collect multiple sensor data quickly
• Data Mining
– POS Analysis required adhoc database queries
Approach
• Leverage the parallelism available in systems
with many disks
• Operate with a small amount of state,
processing data as it streams off the disk
• Execute relatively few instructions per byte of
data
Results- Nearest Neighbor Search
• Problem: Determine k items closest to a
particular item in a database
– Perform comparisons on the drive
– Returns the disks closest matches
– Server does
final merge
Media Mining Example
• Perform low level image
tasks on the disk!
• Edge Detection
performed on disk
– Sent to server as edge
image
– Server does higher level
processing
Why not just use a bunch of PC’s?
• The performance incease is similar
• In fact, the paper essentially used this setup to
actually benchmark their results!
• Supposedly this could be cheaper
• The paper doesn’t really give a good argument
for this…
– Possibly reduced bandwidth on disk IO channel
– But who cares?
Some Questions
• What could a disk possibly do better than the
host processor?
• What added cost is associated with this
mediocre processor on the HDD?
• Are new dependencies are introduced on
hardware and software?
• Perhaps other (better) places to do this type
of local parallel processing?
• Maybe in 2001 this made more sense?